A Voice Dictation System for a Million-Word Czech Vocabulary

The paper describes a set of techniques developed for discrete dictation within a vocabulary that contains up to a million entries, which is one of the main challenges in highly inflected languages like Czech. We present our approach to building an efficiently coded tree lexicon with suffix sub-trees and morphologic classification. Acoustic modeling is based on either monophone, diphone, or triphone models. Lexical and grammatical constraints are represented by unigrams and bigrams, where the latter have form of binary matrix describing grammatical admissibility between word categories. The complete system has been evaluated on a 792,338 word lexicon. Its real-time implementation yields a word-error rate smaller than 12%.